Most electronics manufacturers do not have an AI problem. They have a datasheet problem, and ChatGPT merely makes it visible. The specifications sit in PDFs, customer portals and sales folders. Some of them were never public. The rest are locked in a layout from which AI assistants can only extract the values imperfectly.
Imagine, for example, a design engineer typing into ChatGPT: “Which DIN rail power supply with 24 V and 10 A is suitable for ambient temperatures up to 70 °C?” The answer names three brands, and yours is missing, even though your datasheet answers the question. According to the 2026 State of Marketing to Engineers Report, European technical buyers say they complete almost two thirds of their buying process online before contacting a supplier. That pre-selection happens without you.
How engineers research today
Of the technical buyers who see AI summaries above the search results, only 6% consider them “usually sufficient”. 69% read them and then keep scrolling through the results.
The figures come from the 2026 State of Marketing to Engineers Report by GlobalSpec and TREW Marketing, produced in collaboration with Elektor. More than 1,000 engineers and technical professionals worldwide were surveyed; depending on the question, around 570 to 790 of them answered. Elektor presented the report in German in March 2026.
Even so, generative AI is being used more and more often in the buying process. A good two thirds of respondents use it, although just under a third do so only rarely. 31% still say “never”, down from 42% the previous year. Trust in AI answers stands at 4.7 out of 10 points. At 4.3, Europeans are the biggest sceptics in the regional comparison.
The report itself names one limitation: over 72% of respondents are older than 56, and only 6% are younger than 35. The values therefore mainly describe experienced buyers; younger respondents trust AI answers somewhat more (4.9 points). That does not make things any easier for you, because it is precisely these buyers who double-check an AI answer.
For engineers, AI is therefore the starting point, not the decision. The assistant narrows down the field, and the checking then happens at sources the engineer trusts: trade media, the datasheet, the manufacturer’s website. For Generative Engine Optimisation in B2B, GEO for short, this has two consequences. If your brand is missing from the AI answer, you don’t even make it into this check. If it is in there but your website does not back up the claim, you lose out when the buyer verifies it.
Self-test: what AI assistants say about your product category
Before you optimise, you need to know where you stand, and that takes an hour. Take the ten questions your application engineers hear most often on the phone. Put them to ChatGPT, Perplexity and Gemini the way a design engineer would type them. The DIN rail question from the introduction is the template: specific values, one operating condition, no brand.
For each answer, note three things. Is your brand named? Is one of your pages linked or cited as a source? And who appears instead? The third column is the most interesting. If the AI names a distributor that describes your product, it knows you, just not via your own site. If it only names competitors, it has not drawn on any source about you for this question.
One run is not enough. Rand Fishkin of SparkToro calculates in the same report that across a hundred repetitions of the same ranking question, the chance of ChatGPT or Google’s AI delivering the same list of brands twice is less than one in a hundred. The same question in the morning and in the evening can produce two different lists of manufacturers. Five runs over two weeks tell you more than one. And a screenshot is not a KPI.
If your brand does not appear in a single answer, check the door first. One line in robots.txt that locks out OAI-SearchBot is enough: according to OpenAI’s documentation on its crawlers, your site will then no longer appear in ChatGPT search answers, except at most as a navigation link.
OpenAI documents several crawlers, and three of them matter for visibility in ChatGPT. GPTBot collects training data, OAI-SearchBot feeds ChatGPT search, and ChatGPT-User fetches pages when a user triggers it. If you only block GPTBot, you prevent training and remain findable. Anthropic splits it in a similar way: ClaudeBot collects training data, while Claude-SearchBot indexes for search. What blocks in the firewall and bot management look like, and how to track them down with a curl test as a quick diagnosis, is shown in our article Can AI even find your website?.
How to turn the test questions into a repeatable measurement process is explained in our guide to measuring visibility in ChatGPT, Perplexity and Google AI Overviews. It uses the AI report in Search Console and GA4 for this. For the budget round, a single sentence is then enough, such as “For eight out of ten application questions, the AI names our competitor, and us in none of them.”
Datasheets, product data and application questions
The datasheet for the power supply from the introduction shows the problem. The derating curve is a graphic on one page of the PDF, the temperature limits are in a table with merged cells, and the approvals are logos in the footer. An engineer reads that in seconds. An AI assistant first has to extract it from the layout, and relationships are easily lost in the process. Which value belongs to which variant, which footnote to which row?
That is why the key specifications belong on the product page as an HTML table: rated voltage, output current, temperature range with derating limit, IP rating, standards, form factor. The PDF stays alongside as a download. Google does index PDFs, but what matters is whether every value can be unambiguously assigned to the right product.
The same principle applies to product data in machine-readable format. Manufacturers that classify their items according to ETIM or ECLASS and supply them to wholesalers via BMEcat have long since structured the attributes, from output current to IP rating. If their own product page then only carries a promotional sentence, the gap is plain to see. Add schema.org markup for product and organisation to the attributes there, so that crawlers can unambiguously identify product name, manufacturer and properties.
As of October 2026, Google explicitly states in its documentation on AI features in Search that AI Overviews and AI Mode have no additional technical requirements and need no special markup. The normal rules apply: indexable, crawlable, structured data that matches the visible text. Anyone selling you “GEO markup” as a prerequisite for Google is selling something Google itself does not require.
The ten test questions from the self-test come from your application engineers’ phone calls. They were answered long ago, just by email and each time for a single customer. Your largest untapped resource therefore lies in application engineering, with answers on derating above 50 °C, parallel operation, EMC filters or the replacement for a discontinued model.
Put each of these questions on a page as a subheading and answer it underneath in a short, self-contained paragraph. In this form, an AI assistant can cite it. Under the subheading “Can I connect two 24 V power supplies in parallel?”, for example, there would then be three sentences on whether the model is designed for this, what the user needs to bear in mind and which accessories they need. Add to that a reference to the relevant section of the datasheet.
Trade media: trust is built outside your website
In the report, online trade publications have overtaken manufacturer websites as the most important research source. 76% of technical buyers use them routinely, compared with 74% for manufacturer websites and 21% for generative AI platforms. CJ Haight of GlobalSpec concludes from this that your buyers may be forming their opinion in these publications before they even visit your site. The gap is narrow, as Haight himself admits. For your planning, though, a tie is enough: for your buyers, the trade press counts as much as your own website.
Then there is the question of brand. 70% of respondents tend to choose the better-known brand when faced with two technically similar solutions. Among engineers, awareness is built to a large extent where others write about you: in trade media, at distributors, in forums and in conference programmes.
AI assistants rely on sources like these too. Dale Bertrand of Fire & Spark puts it pointedly in the report: AI learns its answers from reviews, forums and community discussions, not from product pages. I see it a little more cautiously, but I share the direction. An assistant is more likely to name you as a reliable manufacturer of DIN rail power supplies if several independent sources back that up.
Suppose the same power supply manufacturer runs an advertorial titled “The innovative partner for your power supply”. At the same time, its head of development writes a technical article, “Designing correctly for derating above 50 °C”. The second text solves a real design problem and therefore stands a better chance of being linked in forums and drawn on as a source by an AI assistant.
The same applies to application notes published by distributors, to conference talk abstracts and to forum answers from your engineers under their real names. If you have so far filed trade media work under image building, you should reassess it: as input for the systems your customers use to draw up their shortlist.
You won’t build up mentions within one quarter, though. Work on datasheets and application questions pays off faster, which is why it comes first.
GEO check for B2B electronics manufacturers in 5 steps
The first pass fits into an afternoon; repeating the test questions takes two weeks.
- Crawler access (IT or web hosting). Check robots.txt for blocks on OAI-SearchBot, Claude-SearchBot and PerplexityBot, as well as the bot settings of your firewall or CDN. A curl request with a borrowed user agent is only an indication, because bot management recognises real crawlers by IP or signature; the server logs are more reliable. Remove immediately any block on a search crawler that nobody decided on deliberately.
- Test questions (marketing). The ten questions from the self-test, three assistants, five runs over two weeks, results entered in the grid. My threshold: if the AI names a competitor and not you for more than half of the questions, the topic belongs in front of senior management.
- Product pages (product management with the web team). Take your five best-selling products. Is every core specification on the page as text? Is there product markup, and does it match the visible content? If more than two core specifications are missing as text, your bottleneck is the product pages.
- Application questions (application engineering and inside sales). Collect the twenty most frequent questions from the past twelve months and sort them by frequency. Application engineering writes up the first five as answer pages next month.
- Mentions (marketing or PR). Search for your brand name together with the product category in trade media, distributor portals and forums. If most of the hits come from you, trade media work is your bottleneck for next year.
Steps 1 and 3 are the fastest levers, because both are entirely in your own hands. If you want support with implementation or a structured audit, we take that on as a GEO agency, with an audit, technical set-up for crawlers and schema, and content. For the first steps, though, you don’t need us.
Conclusion: application engineering is your real GEO department
The knowledge AI assistants would need about your products lies in application engineering and in the datasheets. For you, GEO mainly means getting this knowledge out of emails and PDFs and onto public pages and into trade media. Marketing organises it; the content comes from engineering.
What I would not recommend: putting budget into “AI-optimised copy” before the search crawlers can get in and the specifications are on the product pages as text. Nor would I stop at your own website. You only earn trust in trade media and forums through technical explanatory work.
According to the report’s authors, the online share of the buying process is rising every year; for European technical buyers it stood at 64% in March 2026. When the engineer from the example asks their 70 °C question again in a few months, the answer should cite your derating table. The first step towards that: open your robots.txt.